The AI boom just took out a mortgage

This week Morgan Stanley put a number on something the bond desks have been muttering about for months: debt issuance tied to artificial intelligence is on track to reach nearly $570 billion this year, more than double last year's total. Through the end of May, AI-linked borrowers had already sold roughly $236 billion of debt globally — about four times what they raised over the same stretch of last year. Let me do what I always do and separate the real thing from the trade built on top of it. The data centers are real. The demand is real. Nothing about this forecast says otherwise, and that was never the question. The question is that for three years the AI boom was an equity story — paid for out of the deepest cash flows in corporate history — and somewhere in the last twelve months it quietly became a credit story. When a boom changes how it is financed, it does not just change its math. It changes its victims.

My father came home from the 2008 crash a quieter man, and the most useful thing he taught me is that the market is a machine for converting other people's certainty into your loss. The certainty in this cycle has been, until now, mostly self-funded certainty. Microsoft, Alphabet, Amazon and Meta built the first leg of this buildout out of their own operating cash flow, which made it — whatever you thought of the prices — the most conservatively financed capex binge in memory. Shareholders rode it, willingly. That era is ending on a schedule you can read in the filings: hyperscaler capital spending this year is on pace to consume close to 100% of operating cash flows, against a ten-year average of around 40%, and Morgan Stanley sees the industry's capex crossing $1 trillion in 2027. You cannot spend more cash than you make and still call it self-funded. So they have sent for the bond desk.

From cash flow to coupon

The turn shows up everywhere once you look for it. Last year the five big hyperscaler names — Amazon, Alphabet, Meta, Microsoft, Oracle — sold about $121 billion of US corporate bonds, against a 2020–2024 average of $28 billion a year. Alphabet, a company that for two decades treated debt as something that happened to other people, is reported to have sold a 100-year bond in February as part of an $85 billion raise. Amazon is reported to have set records in two different currencies in the past year — the largest euro corporate bond sale ever, at €14.5 billion, and the biggest Canadian-dollar 'maple' deal on record. By last October, AI-linked debt had reportedly grown into a $1.2 trillion pile — large enough to overtake US banks as the single biggest sector in JPMorgan's investment-grade index. Read that one again. The largest concentration of blue-chip credit in America is no longer the banking system. It is the buildout.

None of this is irrational, and I want to be precise about that, because every step in a credit cycle is locally rational right up to the step that isn't. If your capex budget has outgrown your cash flow, issuing investment-grade paper at today's spreads is cheaper than selling stock, and the rating agencies will smile on you for a long time before they frown — though S&P is already reported to be warning that Amazon's leverage 'will increase substantially,' with free operating cash flow likely negative for the next two years. The companies are doing the sensible thing, individually. Credit manias are made entirely of companies doing the sensible thing individually.

The precedent everyone has agreed to forget

Here is the history the consensus has filed away, because it is less fun than pets.com. There were two busts at the turn of the century, not one. The dot-com bust — the one we tell stories about — was an equity bust. It vaporized shareholders, who had volunteered. The telecom bust that ran alongside it was a debt bust: fiber networks built on roughly a trillion dollars of borrowing, and when the traffic projections slipped, the result was not sad shareholders but WorldCom and Global Crossing in bankruptcy court, banks eating loan books, and pension funds discovering what they actually owned. The fiber was real — it is still in the ground, still carrying your packets. The debt was the problem. And if you want the longer memory: the railroads of the 1880s, the last time American markets were this concentrated in one story, were built on bonds too. The 1890s were not kind to the people who held them.

An equity bubble burns the people who chose to believe. A debt bubble sends the bill to people who never knew they were at the table.

The reason the distinction matters is that equity and debt fail differently. Equity absorbs disappointment continuously — the price falls, the believers hurt, life goes on. Debt does not absorb disappointment; it schedules it. There are coupons, and there are maturities, and a revenue projection that arrives eighteen months late converts, on a date printed in the indenture, into a default. Equity asks 'what might this be worth?' Credit asks 'can I get my money back?' — and its entire upside is par. A boom funded by believers can deflate. A boom funded by lenders can detonate.

Same boom, very different credits

Now the discipline, because the lazy take here is to see $570 billion and cry havoc, and that is how you lose money in both directions. This debt is not one thing, any more than 'AI' was ever one trade. A bad company and a bad price are different claims; so are a sound borrower and a sound structure. Microsoft — which has notably stayed out of the recent issuance party — Alphabet and Meta could service this paper out of petty cash even if the AI revenue arrived a decade late. Lending to a fortress is boring, and boring is the point.

Then there is Oracle, which one widely read analysis describes as borrowing like a hyperscaler without the balance sheet of one. The company carries a reported $100 billion of total debt against roughly $64 billion in annual revenue; its debt-to-equity ratio is reported around 500%, against something like 23% at Microsoft and Amazon. Its credit default swaps have reportedly traded above 125 basis points — territory last seen across the market in 2009 — which is the market's way of saying it doesn't fully believe the investment-grade rating on the label. And a reported third of Oracle's projected 2028 revenue traces to a single customer, OpenAI, whose own funding needs are the largest in the history of private capital. None of this makes Oracle a bad company. It makes Oracle a leveraged bet on its biggest customer's fundraising calendar, rated as if it were a utility.

Keep going down the stack and the coupons start telling you the truth. xAI is reported to have raised $5 billion at a 12.5% coupon — a number you pay when the lenders have done the math on the collateral. CoreWeave funds itself partly with convertibles. Meta's giant Louisiana data center is financed through a special-purpose vehicle — a reported $27 billion of debt arranged off Meta's balance sheet with PIMCO and Blue Owl — which keeps the parent's credit rating pristine by the simple expedient of putting the leverage somewhere else. Private credit funds already hold a reported $200 billion-plus of this paper, with projections of $300 to $600 billion by 2030, and securitized data-center debt is expected to run $30–40 billion a year. Bank of America's head of leveraged finance, Matt McQueen, said the numbers are 'like nothing any of us who have been in this business for 25 years have seen.' Bankers do not usually advertise that the thing they are selling has no precedent. It is a measure of the moment that it works.

Where the risk goes when it leaves the balance sheet

Risk that leaves a balance sheet does not retire. It migrates — into insurance companies writing fifteen-year private placements against single-tenant buildings, into pension allocations to private credit funds whose marks are a quarterly act of faith, into structured products bought by people three steps removed from the question of what a four-year-old GPU is worth. That last question deserves more respect than it gets: a meaningful slice of this lending is secured, directly or economically, against chips on a two-to-three-year product cadence and buildings designed for exactly one kind of tenant. I wrote earlier this spring about the depreciation schedules; the same argument applies with more force when the asset is collateral instead of a line item. The Bank for International Settlements has said it plainly: AI firms 'now face higher leverage, which could amplify shocks.' And the lender who put it best was Andrew Kleeman at SLC Management — financing data centers right now 'is like selling beer to sailors,' and historically such episodes feature 'massive overinvestment, then correction.' He is still lending, mind you. Everyone is. That is what the middle of a credit cycle looks like from inside.

• Watch spreads, not stories. AI-linked investment-grade paper trading wider than the index — or CDS detaching from a rating, as reportedly happened with Oracle — is the market repricing a narrative before the narrative admits anything.
• Read the footnotes, not the headlines. Special-purpose vehicles, residual-value guarantees and take-or-pay contracts are where the leverage actually lives. Lucent's problem sat in the footnotes for a year before the price noticed; the modern equivalents are better lawyered, not absent.
• Mind the maturity wall. Debt converts disappointment into deadlines. The interesting year is not the year the bonds are sold; it is the year the first big slug of this paper has to roll over into whatever spreads prevail then.
• Ask who holds it. The shift from bank balance sheets to insurers, pensions and private credit means the next stress shows up first where the marks are slowest — and where the holders never thought of themselves as making a bet on AI at all.

What is actually priced in

A word on the round numbers, because they have arrived pre-rounded as usual. $570 billion this year. A trillion of capex in 2027. The buildout itself: $3 trillion by Morgan Stanley's arithmetic, more than $5 trillion by JPMorgan's. When two of the most sophisticated credit shops on earth disagree by two trillion dollars about the size of the thing they are both underwriting, the disagreement is the most honest data point in either report. And I will pay my usual toll: I have called two bubbles correctly and one about eighteen months early, and being early is a way of being wrong that cost real people real money while I felt clever. So I am not calling the top. The borrowing can run for years. Credit cycles are long precisely because everyone in them is being individually reasonable.

But I can tell you what today's spreads have already agreed to believe. They believe the AI revenue arrives roughly on schedule. They believe the chips hold their value as long as the depreciation tables claim. They believe a half-trillion dollars a year of new paper can be rolled, on dates already printed, at prices not much worse than today's. Each of those can be true. None of them is free, and the coupon you are being paid does not price all three. The fiber was real, and the bondholders still lost. The question I ask of every boom is the one my father taught me — compared to when? This one, increasingly, compares to the telecoms: a real technology, an honest demand curve, and a capital structure quietly trading its shock absorbers for deadlines. The buildout used to be the shareholders' bet. As of this year, it is becoming everyone's. Watch the spread, not the story.